Back to Blog
Insight2026-03-26 · 14 min read

Claude Takes the Wheel on Your Mac: How Dispatch and Computer Use Redraw Agent Boundaries

In March 2026 Anthropic shipped Computer Use in Claude Cowork and Claude Code alongside Dispatch. This post ties official docs, press, and hands-on reviews into how it works, what it requires, and where the risks sit.

Claude Takes the Wheel on Your Mac: How Dispatch and Computer Use Redraw Agent Boundaries

The next frontier for agentic AI is not talking but acting on your tools. In March 2026 Anthropic added Computer Use to Claude Cowork and Claude Code as a research preview, and paired it with Dispatch so you can give instructions from your phone and see the finished work on your Mac.

This article draws on the Anthropic blog, Help Center, MacRumors, VentureBeat, and MacStories’ Dispatch hands-on to explain what changed, how tool selection is ordered, and what to assume on safety and governance.

In one sentence

  • Computer Use: Inside Cowork and Code, Claude can finish tasks by driving mouse, keyboard, and screen across apps and browsers—not a separate API surface, and framed as low-setup.
  • Dispatch: A continuous conversation pipeline where you assign work from mobile (e.g. iPhone) and execution completes on Mac, with the desktop app awake.
  • Current limits: macOS only, Claude Pro / Max, research preview—for complex multi-app jobs, keep expectations modest on reliability and speed.

Why “screen control” now

Beyond summarizing documents, people and teams want unbroken workflows across Slack, calendar, browser, IDE. Anthropic’s post says that when connectors (Slack, Google Calendar, etc.) are missing, Claude can explore the screen directly. The hierarchy is explicit: best connector → browser → GUI as last resort.

Press such as VentureBeat frames this as competition to build agents that actually work, alongside tension with open-source computer-automation stacks. (This blog covers OSS tools too; here the focus is design trade-offs in a commercial product.)

What Computer Use does (official framing)

Per the official post (2026-03-23):

  • When the right tool is not wired in, Claude can navigate with pointer, clicks, and scroll to complete the task.
  • It can try opening files, using the browser, and running dev tools with minimal extra setup, per Anthropic.
  • Explicit user approval is required before accessing a new app.
  • To mitigate prompt injection, model activation during computer use is automatically scanned, they say.
  • They also state plainly that computer use is still early, mistakes happen, and threats evolve. They recommend trusted apps first and avoiding sensitive data.

That mix is both marketing and safety copy; as a user, it helps to separate convenience from who bears responsibility when something goes wrong.

Tool order: why the screen comes last

Combining VentureBeat and Help Center material, the rough order is:

  1. First-party connectors (Gmail, Slack, Google Calendar, Drive, etc.)—fast and stable.
  2. Browser paths such as Claude for Chrome—the middle layer when no connector exists.
  3. Screenshot-driven GUI control—maximum generality, but slower and more failure-prone.

Help Center copy compares fetching Slack via connector (seconds) with driving Slack through the UI (much slower, higher error rate). That is product philosophy and engineering reality. Screen control is not a master key; it is a detour.

Dispatch: your phone as a remote

Dispatch landed in Cowork first; coverage describes extension to Claude Code in the same wave. After pairing mobile and Mac (e.g. via QR), you assign work in natural language from the phone and Claude runs on the Mac.

Official examples are intuitive:

  • A morning briefing while commuting
  • Edit, test, and open a PR from the IDE
  • Keep a long-running task like 3D printing aligned with an initial plan mid-stream

The prerequisites are clear: the desktop app must be awake and running, and only macOS is supported. Marketing often says “your Mac works while you’re away,” but in practice your Mac is always the session host.

Pricing, platform, release status

  • Claude Pro / Max subscribers (pricing varies by region and tier; press often cites Pro around $17/month and higher Max tiers).
  • macOS-only research preview; Windows users are out of scope for now.
  • Anthropic states results may not be perfect and complex jobs may need a second attempt.

Security and compliance: read the footnotes with the hype

1) What’s on screen is what the model can “see”

GUI automation relies on screenshots and UI interpretation. VentureBeat notes sensitive on-screen data can be exposed, and that even when Anthropic steers away from certain actions (e.g. trading, sensitive inputs, facial imagery), guardrails are not absolute.

2) Permissions, block lists, injection scanning

Layers include per-app allowlists, default blocks for some sensitive categories (investing, crypto, etc., per reporting), user blocklists, and stop controls.

Still, mistakes, malicious pages, and malicious instructions remain plausible. Because the agent path is close to keyboard and mouse as the user, audit and forensics questions follow quickly. VentureBeat highlights a gap between enterprise features and centralized audit of Cowork activity.

3) Real-world difficulty

MacStories’ John Voorhees reports that for Dispatch-related flows, research and summaries work, but across attempts success and failure were closer to fifty-fifty—aligned with the “research preview” label. Plan for demo vs. production gap.

Takeaways for developers and architects

  1. Integration priority: When designing automation, APIs, events, and official connectors still win. Screen control is a bridge for legacy or closed UIs.
  2. Observability: For team use, do not rely only on product defaults for who touched which app, when—define policy and logging explicitly.
  3. Prompt-injection surface: Browser, document, and chat content become agent input. Refresh threat models assuming classic web risks can spill into desktop actions.
  4. Cost and quotas: Agentic runs burn tokens and plan limits quickly via sub-agents and tool calls (per user reports and press). For internal PoCs, design budget caps up front.

Closing

Computer Use + Dispatch pushes the vision of Claude as an operator on your Mac, not just a chat window. Anthropic also draws a line: accuracy, speed, and security are still early.

If you evaluate adoption, three practices help:

  • Scope to trusted apps and data.
  • Prefer connectors and APIs wherever they exist.
  • Read “research preview” as no SLA.

References

Written from public reporting, documentation, and reviews. Pricing and feature availability can change; confirm the latest announcements and your account settings before relying on them in production.